Reproducing DeepTFUS | projects

Focused ultrasound can reach targets deep inside the brain non-invasively, but figuring out where the beam actually lands through someone's skull takes ~30 minutes of physics simulation per patient. I reproduced [DeepTFUS](https://arxiv.org/abs/2505.12998), a 3D model that **predicts the same result in under a second**, and open-sourced the trained weights + training code the paper didn't release.

DeepTFUS is a proposed model that, given a CT + ultrasound transducer placement, predicts the resulting pressure field inside the head in <1s instead of the ~30m a physics simulator would take. Model weights and training code are not released, so I: open-source my reproduction , walk through some new loss functions I tried , and poke around with the trained model Update: My reproduction has 11x less parameters (3.4M vs ~38M; post written before the authors got back with architecture details, now in the Update ). This was a 36-hour learning sprint on an H100, no sweeps or anything rigorous. gro

saved by

• Cheikh Fiteni
• agniv sarkar

related reading

• (Some of) The Models, They Just Don't Want to Learn | Tildeblog.tilderesearch.com
• The Annotated Transformernlp.seas.harvard.edu
• Transformers from Scratche2eml.school
• The Decade of Deep Learning | Leo Gaobmk.sh
• Tiny TPUtinytpu.com
• The Little Book of Deep Learningfleuret.org
• The Annotated Transformernlp.seas.harvard.edu
• We Induced Artificial Smells With Ultrasound Brain Stimulation | Write to Brainwritetobrain.com
• Transformer Math 101 | EleutherAI Blogblog.eleuther.ai
• 1706.03762arxiv.org
• GitHub - google-research/tuning_playbook: A playbook for systematically maximizing the performance of deep learning models. · GitHubgithub.com
• Full-waveform inversion imaging of the human brain | npj Digital Medicinenature.com